[GitHub] [tvm] Mousius commented on a change in pull request #8951: [3/10] Moved TIR generation from Python to C++ for CMSIS-NN

[GitBox <gi...@apache.org>]

Mousius commented on a change in pull request #8951:
URL: https://github.com/apache/tvm/pull/8951#discussion_r703699778



##########
File path: python/tvm/relay/backend/contrib/cmsisnn/codegen.py
##########
@@ -0,0 +1,34 @@
+# Licensed to the Apache Software Foundation (ASF) under one
+# or more contributor license agreements.  See the NOTICE file
+# distributed with this work for additional information
+# regarding copyright ownership.  The ASF licenses this file
+# to you under the Apache License, Version 2.0 (the
+# "License"); you may not use this file except in compliance
+# with the License.  You may obtain a copy of the License at
+#
+#   http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, either express or implied.  See the License for the
+# specific language governing permissions and limitations
+# under the License.
+"""Codegen for CMSIS-NN"""
+import tvm
+
+
+@tvm.register_func("relay.ext.cmsisnn")
+def cmsisnn_compiler(relay_func):

Review comment:
       Is there any reason to keep this in Python? I think this can be registered in C++ and the entire set of Python files removed now?

##########
File path: src/relay/backend/contrib/cmsisnn/relay_to_tir.cc
##########
@@ -0,0 +1,147 @@
+
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one
+ * or more contributor license agreements.  See the NOTICE file
+ * distributed with this work for additional information
+ * regarding copyright ownership.  The ASF licenses this file
+ * to you under the Apache License, Version 2.0 (the
+ * "License"); you may not use this file except in compliance
+ * with the License.  You may obtain a copy of the License at
+ *
+ *   http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+ * KIND, either express or implied.  See the License for the
+ * specific language governing permissions and limitations
+ * under the License.
+ */
+#include <tvm/relay/expr_functor.h>
+#include <tvm/tir/builtin.h>
+#include <tvm/tir/expr.h>
+#include <tvm/tir/function.h>
+#include <tvm/tir/op.h>
+#include <tvm/tir/stmt_functor.h>
+
+#include "../../../qnn/utils.h"
+
+namespace tvm {
+namespace relay {
+
+class RelayToTIR : public MixedModeVisitor {
+ public:
+  explicit RelayToTIR(String func_name) : func_name_(func_name) {}
+
+ private:
+  bool is_quant_softmax(const CallNode* call) {
+    const OpNode* op = call->op.as<OpNode>();
+    if (op == nullptr || op->name != "qnn.quantize") {
+      return false;
+    }
+    const CallNode* softmax = call->args[0].as<CallNode>();
+    op = softmax->op.as<OpNode>();
+    if (op->name != "nn.softmax") {
+      return false;
+    }
+    const CallNode* dequantize = softmax->args[0].as<CallNode>();
+    op = dequantize->op.as<OpNode>();
+    if (op->name != "qnn.dequantize") {
+      return false;
+    }
+    return true;
+  }
+
+  void emit_softmax_tir(const CallNode* call) {
+    auto* softmax_call = call->args[0].as<CallNode>();
+    auto* dequant_call = softmax_call->args[0].as<CallNode>();
+    auto* scale_const = dequant_call->args[1].as<ConstantNode>();
+    const float quant_scale = static_cast<const float*>(scale_const->data->data)[0];
+
+    // assuming layout as NHWC
+    auto shape = call->type_as<TensorTypeNode>()->shape;
+    int trailing_dim = shape.size() - 1;
+    int row_size = shape[trailing_dim].as<tir::IntImmNode>()->value;
+    int num_rows = 1;
+    for (int i = 0; i < trailing_dim; ++i) {
+      num_rows *= shape[i].as<tir::IntImmNode>()->value;
+    }
+
+    // calculate multiplier and shift for CMSIS-NN softmax API
+    // Note: TensorFlow Lite Micro assumptions
+    // Output zero point and scale are fixed to -128 and 1 / 256
+    double beta = 1.0;
+    int32_t input_bits = 5;
+    double beta_multiplier = (beta * quant_scale * (1 << (31 - input_bits)));
+    beta_multiplier = std::min<double>(beta_multiplier, (1ll << 31) - 1.0);
+    auto mult_shift_pair = tvm::relay::qnn::GetFixedPointMultiplierShift(beta_multiplier);
+    int32_t mult = std::get<0>(mult_shift_pair);
+    int32_t shift = std::get<1>(mult_shift_pair);
+    int32_t diff_min = (1 << 5) - 1;
+    diff_min <<= (31 - 5);
+    diff_min >>= shift;
+    diff_min *= -1;
+
+    auto in_var = tir::Var("input", DataType::Handle(8));
+    auto out_var = tir::Var("output", DataType::Handle(8));
+
+    Array<tir::Var> main_signature{in_var, out_var};
+
+    tvm::Array<PrimExpr> args;
+    args.push_back(tir::StringImm("arm_softmax_s8"));
+    args.push_back(in_var);
+    args.push_back(IntImm(DataType::Int(32), num_rows));
+    args.push_back(IntImm(DataType::Int(32), row_size));
+    args.push_back(IntImm(DataType::Int(32), mult));
+    args.push_back(IntImm(DataType::Int(32), shift));
+    args.push_back(IntImm(DataType::Int(32), diff_min));
+    args.push_back(out_var);

Review comment:
       Does this not work initialised in one go? For example:
   ```suggestion
       tvm::Array<PrimExpr> args = {
           tir::StringImm("arm_softmax_s8"),
           in_var,
           IntImm(DataType::Int(32), num_rows),
           IntImm(DataType::Int(32), row_size),
           IntImm(DataType::Int(32), mult),
           IntImm(DataType::Int(32), shift),
           IntImm(DataType::Int(32), diff_min),
           out_var
       };
   ```

##########
File path: src/relay/backend/contrib/cmsisnn/codegen_cmsisnn.cc
##########
@@ -0,0 +1,190 @@
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one
+ * or more contributor license agreements.  See the NOTICE file
+ * distributed with this work for additional information
+ * regarding copyright ownership.  The ASF licenses this file
+ * to you under the Apache License, Version 2.0 (the
+ * "License"); you may not use this file except in compliance
+ * with the License.  You may obtain a copy of the License at
+ *
+ *   http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+ * KIND, either express or implied.  See the License for the
+ * specific language governing permissions and limitations
+ * under the License.
+ */
+#include <cmath>
+#include <fstream>
+#include <map>
+#include <sstream>
+#include <string>
+#include <vector>
+
+#include "../../../../runtime/file_utils.h"
+#include "../../../../target/source/codegen_c.h"
+
+namespace tvm {
+namespace runtime {
+
+using namespace tir;
+
+class CodeGenCMSISNN : public tvm::codegen::CodeGenC {
+ public:
+  void Init(bool output_ssa) {
+    decl_stream << "#include <stdio.h>\n";
+    decl_stream << "#include <stdlib.h>\n";
+    decl_stream << "#include <dlpack/dlpack.h>\n";
+    decl_stream << "#include <tvm/runtime/crt/module.h>\n";
+    decl_stream << "#include <arm_nnfunctions.h>\n";
+    CodeGenC::Init(output_ssa);
+  }
+
+  /*!
+   * \brief Emit code that offloads a subgraph to the Cortex-M
+   *
+   * \return string of code that offloads a subgraph to the Cortex-M
+   */
+  void AddFunction(const PrimFunc& prim_func) {
+    PrintExternCPrefix(stream);
+    CodeGenC::AddFunction(prim_func);
+    PrintExternCPostfix(stream);
+  }
+
+ private:
+  void VisitExpr_(const CallNode* op, std::ostream& os) {  // NOLINT(*)
+    if (!op->op.same_as(builtin::call_extern())) {
+      return;
+    }
+    std::string cmsis_func_name = op->args[0].as<StringImmNode>()->value;
+    if (cmsis_func_name == "arm_softmax_s8") {
+      EmitSoftmax(op);

Review comment:
       Can we not use the default behaviour now for `arm_softmax_s8` due to it emitting a standard `call_extern`? I'd still leave the infrastructure in for the struct calls but I don't think we need it specifically for this call?

##########
File path: src/relay/backend/contrib/cmsisnn/relay_to_tir.cc
##########
@@ -0,0 +1,147 @@
+
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one
+ * or more contributor license agreements.  See the NOTICE file
+ * distributed with this work for additional information
+ * regarding copyright ownership.  The ASF licenses this file
+ * to you under the Apache License, Version 2.0 (the
+ * "License"); you may not use this file except in compliance
+ * with the License.  You may obtain a copy of the License at
+ *
+ *   http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+ * KIND, either express or implied.  See the License for the
+ * specific language governing permissions and limitations
+ * under the License.
+ */
+#include <tvm/relay/expr_functor.h>
+#include <tvm/tir/builtin.h>
+#include <tvm/tir/expr.h>
+#include <tvm/tir/function.h>
+#include <tvm/tir/op.h>
+#include <tvm/tir/stmt_functor.h>
+
+#include "../../../qnn/utils.h"
+
+namespace tvm {
+namespace relay {
+
+class RelayToTIR : public MixedModeVisitor {
+ public:
+  explicit RelayToTIR(String func_name) : func_name_(func_name) {}
+
+ private:
+  bool is_quant_softmax(const CallNode* call) {
+    const OpNode* op = call->op.as<OpNode>();
+    if (op == nullptr || op->name != "qnn.quantize") {
+      return false;
+    }
+    const CallNode* softmax = call->args[0].as<CallNode>();
+    op = softmax->op.as<OpNode>();
+    if (op->name != "nn.softmax") {
+      return false;
+    }
+    const CallNode* dequantize = softmax->args[0].as<CallNode>();
+    op = dequantize->op.as<OpNode>();
+    if (op->name != "qnn.dequantize") {
+      return false;
+    }
+    return true;
+  }
+
+  void emit_softmax_tir(const CallNode* call) {
+    auto* softmax_call = call->args[0].as<CallNode>();
+    auto* dequant_call = softmax_call->args[0].as<CallNode>();
+    auto* scale_const = dequant_call->args[1].as<ConstantNode>();
+    const float quant_scale = static_cast<const float*>(scale_const->data->data)[0];
+
+    // assuming layout as NHWC
+    auto shape = call->type_as<TensorTypeNode>()->shape;
+    int trailing_dim = shape.size() - 1;
+    int row_size = shape[trailing_dim].as<tir::IntImmNode>()->value;
+    int num_rows = 1;
+    for (int i = 0; i < trailing_dim; ++i) {
+      num_rows *= shape[i].as<tir::IntImmNode>()->value;
+    }
+
+    // calculate multiplier and shift for CMSIS-NN softmax API
+    // Note: TensorFlow Lite Micro assumptions
+    // Output zero point and scale are fixed to -128 and 1 / 256
+    double beta = 1.0;
+    int32_t input_bits = 5;
+    double beta_multiplier = (beta * quant_scale * (1 << (31 - input_bits)));
+    beta_multiplier = std::min<double>(beta_multiplier, (1ll << 31) - 1.0);
+    auto mult_shift_pair = tvm::relay::qnn::GetFixedPointMultiplierShift(beta_multiplier);
+    int32_t mult = std::get<0>(mult_shift_pair);
+    int32_t shift = std::get<1>(mult_shift_pair);
+    int32_t diff_min = (1 << 5) - 1;
+    diff_min <<= (31 - 5);
+    diff_min >>= shift;
+    diff_min *= -1;
+
+    auto in_var = tir::Var("input", DataType::Handle(8));
+    auto out_var = tir::Var("output", DataType::Handle(8));
+
+    Array<tir::Var> main_signature{in_var, out_var};
+
+    tvm::Array<PrimExpr> args;
+    args.push_back(tir::StringImm("arm_softmax_s8"));
+    args.push_back(in_var);
+    args.push_back(IntImm(DataType::Int(32), num_rows));
+    args.push_back(IntImm(DataType::Int(32), row_size));
+    args.push_back(IntImm(DataType::Int(32), mult));
+    args.push_back(IntImm(DataType::Int(32), shift));
+    args.push_back(IntImm(DataType::Int(32), diff_min));
+    args.push_back(out_var);
+    tir::Stmt body =
+        tir::Evaluate(tvm::tir::Call(DataType::Int(8), tir::builtin::call_extern(), args));
+
+    Map<String, ObjectRef> dict_attrs;
+    dict_attrs.Set("global_symbol", func_name_);
+    dict_attrs.Set("tir.noalias", Bool(true));
+
+    primfunc_ = tir::PrimFunc(main_signature, body, VoidType(), Map<tir::Var, tir::Buffer>(),
+                              DictAttrs(dict_attrs));
+  }
+
+  void VisitExpr_(const CallNode* call) final {
+    if (is_quant_softmax(call)) {
+      emit_softmax_tir(call);
+    }
+  }
+
+ public:
+  String func_name_;

Review comment:
       As these are public, they shouldn't have `_`s

##########
File path: src/relay/backend/contrib/cmsisnn/relay_to_tir.cc
##########
@@ -0,0 +1,147 @@
+
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one
+ * or more contributor license agreements.  See the NOTICE file
+ * distributed with this work for additional information
+ * regarding copyright ownership.  The ASF licenses this file
+ * to you under the Apache License, Version 2.0 (the
+ * "License"); you may not use this file except in compliance
+ * with the License.  You may obtain a copy of the License at
+ *
+ *   http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+ * KIND, either express or implied.  See the License for the
+ * specific language governing permissions and limitations
+ * under the License.
+ */
+#include <tvm/relay/expr_functor.h>
+#include <tvm/tir/builtin.h>
+#include <tvm/tir/expr.h>
+#include <tvm/tir/function.h>
+#include <tvm/tir/op.h>
+#include <tvm/tir/stmt_functor.h>
+
+#include "../../../qnn/utils.h"
+
+namespace tvm {
+namespace relay {
+
+class RelayToTIR : public MixedModeVisitor {
+ public:
+  explicit RelayToTIR(String func_name) : func_name_(func_name) {}
+
+ private:
+  bool is_quant_softmax(const CallNode* call) {
+    const OpNode* op = call->op.as<OpNode>();
+    if (op == nullptr || op->name != "qnn.quantize") {
+      return false;
+    }
+    const CallNode* softmax = call->args[0].as<CallNode>();
+    op = softmax->op.as<OpNode>();
+    if (op->name != "nn.softmax") {
+      return false;
+    }
+    const CallNode* dequantize = softmax->args[0].as<CallNode>();
+    op = dequantize->op.as<OpNode>();
+    if (op->name != "qnn.dequantize") {
+      return false;
+    }
+    return true;
+  }
+
+  void emit_softmax_tir(const CallNode* call) {
+    auto* softmax_call = call->args[0].as<CallNode>();
+    auto* dequant_call = softmax_call->args[0].as<CallNode>();
+    auto* scale_const = dequant_call->args[1].as<ConstantNode>();
+    const float quant_scale = static_cast<const float*>(scale_const->data->data)[0];
+
+    // assuming layout as NHWC
+    auto shape = call->type_as<TensorTypeNode>()->shape;
+    int trailing_dim = shape.size() - 1;
+    int row_size = shape[trailing_dim].as<tir::IntImmNode>()->value;
+    int num_rows = 1;
+    for (int i = 0; i < trailing_dim; ++i) {
+      num_rows *= shape[i].as<tir::IntImmNode>()->value;
+    }
+
+    // calculate multiplier and shift for CMSIS-NN softmax API
+    // Note: TensorFlow Lite Micro assumptions
+    // Output zero point and scale are fixed to -128 and 1 / 256
+    double beta = 1.0;
+    int32_t input_bits = 5;
+    double beta_multiplier = (beta * quant_scale * (1 << (31 - input_bits)));
+    beta_multiplier = std::min<double>(beta_multiplier, (1ll << 31) - 1.0);
+    auto mult_shift_pair = tvm::relay::qnn::GetFixedPointMultiplierShift(beta_multiplier);
+    int32_t mult = std::get<0>(mult_shift_pair);
+    int32_t shift = std::get<1>(mult_shift_pair);
+    int32_t diff_min = (1 << 5) - 1;
+    diff_min <<= (31 - 5);
+    diff_min >>= shift;
+    diff_min *= -1;
+
+    auto in_var = tir::Var("input", DataType::Handle(8));
+    auto out_var = tir::Var("output", DataType::Handle(8));
+
+    Array<tir::Var> main_signature{in_var, out_var};

Review comment:
       Can we name this to reflect it isn't the main function?




-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: commits-unsubscribe@tvm.apache.org

For queries about this service, please contact Infrastructure at:
users@infra.apache.org